Cheap AI API keys from third parties: what companies risk

heise online bought a grey-market API key for frontier AI models at a fraction of the list price and had it tested by security specialists. The findings matter to any company whose staff might use such access without telling IT.

Klaus Bauer5 min read

Illustration: A key with a discount tag sits by a poorly sealed parcel, next to an open padlock and spilled coins.

What heise online found after buying a cheap key

The technology magazine heise online published an analysis in its paid heise+ section, KI-Zugang für 2 Euro, in November 2026. The editors found a seller on the GamsGo marketplace, a platform for account sharing. Third-party vendors there offer access to frontier models such as Claude Fable 5.1, Claude Opus 5 and OpenAI's GPT-6 Astra. The editors bought access there for a very small sum.

They then handed the key to whitelisthackers, a company for ethical hacking and cybersecurity. The testers ran a pentesting tool and several thousand prompts against it. Anthropic models did respond, but the traces lead through an opaque gateway to a Russian API service and to Amazon's development environment Kiro.

Nobody can verify which model actually answers. The security expert quoted by heise concludes that the operator can read every prompt in plain text and that requests are altered. The risks named are data protection violations, manipulated answers and the loss of the purchased balance. Only a teaser of the article is freely available, so details beyond this summary remain open.

Why the prices are so low and what that means

The starting point is the cost of top models. According to heise, citing the benchmark platform Artificial Analysis, Anthropic's Claude Fable 5.2 leads the intelligence ranking and costs 40 US dollars per million generated tokens. A heavily used workflow can consume that in a day. Some vendors promise discounts of up to 95 percent, which is hard to reach through legitimate channels.

The teaser does not explain how the sellers finance their prices, so that question stays open here. It does show that the key leads to a chain of intermediaries and not directly to the model provider. Each intermediary in that chain sees the traffic. A buyer has no contract with most of them and no way to check what they store or change.

Are such accesses secure and GDPR-compliant?

Based on the heise findings, a company has little basis to assume so. If the operator reads prompts in plain text, every customer name, contract excerpt or employee detail typed into the tool reaches an unknown party. Under the GDPR a company that passes personal data to a service provider normally needs a clear contract, knowledge of who processes the data and a basis for any transfer abroad. A reseller in a grey market offers none of that in a verifiable form.

For holders of professional secrets such as lawyers and tax advisers, section 203 of the German Criminal Code (StGB) adds a criminal-law dimension to unauthorised disclosure. Whether a specific case is lawful needs an assessment by your data protection officer or counsel. The manipulation finding creates a second problem: altered requests and unverifiable models make answers unreliable, even when no personal data is involved.

What companies should check now

Shadow IT means tools that staff use without approval from IT or management. Cheap AI keys spread this way, often through a developer or a department that wants to save budget. A short review usually covers the following points.

  • Ask developers and departments which API keys, browser extensions and AI tools they use, and where the keys came from.
  • Check expense reports and corporate cards for payments to marketplaces and account-sharing platforms.
  • Search code repositories and configuration files for API keys and for unusual endpoint addresses.
  • Block unknown AI endpoints at the proxy or firewall where this is practical, and rotate keys that may have been exposed.
  • Write a short rule on approved AI services and tell staff who to ask for a new tool.
  • Involve the Betriebsrat (works council) where monitoring of usage is planned, and your data protection officer for any incident.

If a key has already been used with real company data, treat it as a possible data leak and document what was entered. The guidance on GDPR-compliant AI and the page on security and data protection outline the questions a provider should be able to answer. Mittelstand firms, the German small and medium-sized companies, can start with the approach in introducing AI in SMEs.

What a regulated access can look like

A regulated route gives staff a sanctioned tool, so nobody needs to look for a bargain key. The provider should be named, the hosting location known and the contract clear. Costs should be predictable, because unpredictable token bills push teams towards cheap workarounds.

souveraen.ai, from clouditiv GmbH in Auerbach, is one such platform for company knowledge. It runs On Demand, hosted in Germany, or on-premise as Private Spark on an NVIDIA DGX Spark at a fixed monthly price without token costs for local models. The model catalogue covers local models, models hosted in Germany and external EU models with zero data retention. Administrators get roles, workspaces and a usage overview, which makes visible who uses what.

Whatever platform you choose, ask for the same things: a named operator, documented hosting, a contract and a way to see usage. The developer API of souveraen.ai is OpenAI-compatible, which is the interface many developer tools already expect.

Sources

More articles

All articles

How safe are AI API routes via resellers and routers?

Security firm Dragos sent about two million requests with fake credentials to AI providers and found compromised services and resellers reading customer prompts. Companies that use AI through third parties need to know who sits on the path of their data.

Data protection4 min

Run Text Generation Inference with Docker on Your Server

Text Generation Inference (TGI) serves open language models from a single Docker container on your own hardware. This tutorial starts it, sends requests over HTTP and keeps prompts inside your network.

Tutorial4 min

Cloud outage liability: who answers when Microsoft fails

On 12 November 2026 a software fault took Exchange Online and Outlook offline for about 36 hours. A legal analysis published shortly afterwards argues that responsibility for such outages often lies with the operator rather than the provider.

Regulation4 min

Test it with your own documents.

Free for 3 days, then from €24.90 per user per month. Cancellable monthly.

Free for 3 days, then from €24.90 per user per month.Free for 3 days
Try free for 3 days